Wealth Engine

Category: AI News

  • DeepSeek Vision Model vs Claude Opus 4.8: 23x Cheaper Per Image

    DeepSeek’s vision model just split a 2-2 draw with Claude Opus 4.8 on visual benchmarks while charging roughly 23 times less per image. V4-Flash-Vision-Exp processes 1,000 images for about $0.17 at peak rates. Claude Sonnet 4.6 charges around $4 for the same work. If your workload is image-heavy and volume-driven, the math is no longer close.

    The DeepSeek vision model shipped on August 21, 2026. It is experimental, API-only, and it does something DeepSeek has done before: it removes a price floor the incumbents were standing on.

    What is DeepSeek’s new vision model?

    V4-Flash-Vision-Exp is an experimental multimodal extension of DeepSeek’s V4-Flash, adding image understanding to a model that was previously text-only. It is available on DeepSeek’s paid developer platform by setting model='deepseek-v4-flash-vision-exp', per the official API changelog.

    The target is not image generation. It is visual agent work — reading charts, parsing screenshots, driving interfaces, and running multi-step tasks that require seeing what happened.

    DeepSeek’s own framing is careful. The changelog calls it “a significant leap over DeepSeek-V4-Flash, bringing its multimodal agent capabilities close to Opus-4.8,” while noting pure-text performance is unchanged from the base model.

    What’s under the hood

    The foundation is V4-Flash, released in April 2026: a 284-billion-parameter mixture-of-experts model built from multiple 13-billion-parameter networks, according to SiliconANGLE.

    The vision specs are aggressive. A 1-million-token context window, up to 384,000 output tokens, and up to 600 images per request. Each image is capped at 384 tokens regardless of size — that cap is the entire reason the pricing works.

    Supported formats are JPEG, PNG, GIF and WebP, with a maximum edge of 8,192 pixels, dropping to 4,096 pixels once you send 15 or more images in a single call.

    Did the DeepSeek vision model actually beat Claude Opus 4.8?

    On two benchmarks yes, on two no. It is a genuine 2-2 split, not a sweep, and the margins are thin in both directions. DeepSeek wins the agentic and hard-visual tests. Opus 4.8 holds the chart-reading and structured-reasoning tests. Anyone claiming a decisive winner is selling something.

    Benchmark DeepSeek V4-Flash-Vision-Exp Claude Opus 4.8 Margin
    Agents’ Last Exam (ALE) 27.3 25.7 +1.6 DeepSeek
    ZeroBench (Pass@5) 35.0 34.0 +1.0 DeepSeek
    ApexBench (Pass@1) 36.5 39.4 +2.9 Opus 4.8
    Chartography 64.3 65.0 +0.7 Opus 4.8

    Scores as reported by XenoSpectrum’s breakdown. DeepSeek’s changelog separately lists a Terminal Bench 2.1 score of 83.9.

    Context on the tests matters. ALE contains more than 1,000 multi-step tasks; ZeroBench is 100 image-analysis tasks deliberately built to be hard, SiliconANGLE reports. Those are the two DeepSeek won.

    The pattern is consistent with what we saw when GLM-5.3 went up against DeepSeek V4 Pro: Chinese labs are now trading benchmark wins with frontier Western models rather than trailing them.

    How much does the DeepSeek vision model cost?

    Very little, and the structure is unusual. DeepSeek runs peak and off-peak pricing. Off-peak, input is $0.22 per million tokens and output is $0.66. During peak hours — 01:00–04:00 and 06:00–10:00 UTC — both double, to $0.44 and $1.32. Cache hits cost $0.007 off-peak.

    The image economics fall out of the 384-token cap. A single image costs $0.00008448 off-peak and $0.00016896 at peak.

    Model ~Cost per 1,000 images (1024×1024) Input / output per 1M tokens Context
    DeepSeek V4-Flash-Vision-Exp (off-peak) ~$0.09 $0.22 / $0.66 1M
    DeepSeek V4-Flash-Vision-Exp (peak) ~$0.17 $0.44 / $1.32 1M
    Google Gemini 3.1 Pro ~$0.50 $2.00 / $12.00 1M
    OpenAI GPT-5.4 Vision ~$1.90 $2.50 / $15.00 1.1M
    Anthropic Claude Sonnet 4.6 ~$4.00 — —
    Anthropic Claude Opus 4.8 — $5.00 / $25.00 1M

    Per-image costs are as calculated by BigGo Finance; Opus 4.8 token rates are $5.00 input, $25.00 output and $0.50 cached input, per OpenRouter.

    Set those side by side. On output tokens, DeepSeek off-peak at $0.66 against Opus 4.8 at $25.00 is a 37x spread for a model that just won two of four visual benchmarks against it.

    Why the 384-token image cap matters

    Most vision APIs price images by resolution. Send a bigger image, pay more tokens. That makes cost forecasting miserable and punishes exactly the high-resolution documents that need vision in the first place.

    DeepSeek caps every image at 384 tokens no matter its size, up to an 8,192-pixel edge. A dense scanned contract and a low-resolution thumbnail cost the same.

    For finance and operations teams, that turns a variable line item into a fixed one. You can price a document-processing product per page before you have seen a single customer’s files, which is not something the resolution-scaled competitors let you do.

    Is the DeepSeek vision model cheap enough to switch?

    For high-volume image pipelines, yes — the gap is too large to argue with. For low-volume work where quality variance costs more than tokens, no. The deciding number is how many images you process per month, because below roughly 100,000 images the absolute savings stop mattering.

    Run it concretely. A million images per month costs about $170 at DeepSeek’s peak rate, $90 off-peak, and roughly $4,000 on Claude Sonnet 4.6. That is a $46,000 annual difference on a single pipeline.

    • Document and invoice processing at scale. Switch. The 384-token cap makes cost per page nearly flat.
    • Screenshot-driven agents and RPA. Switch. ALE and ZeroBench are exactly this workload, and DeepSeek won both.
    • Chart and dashboard extraction. Stay. Opus 4.8 leads Chartography 65.0 to 64.3, and errors here are expensive.
    • Regulated or air-gapped work. Stay. This is a Chinese paid API with no open weights announced.
    • Batch archive backfill. Switch, and schedule it off-peak. You halve the bill by choosing when to run.

    That off-peak lever is genuinely underrated. Almost no batch job cares what hour it runs, and DeepSeek is effectively paying you 50% to move it.

    Which vision model should you use for what?

    Match the model to the failure cost, not to the benchmark chart. Where a wrong answer is cheap and volume is enormous, take the cheapest capable model. Where a wrong answer triggers a human review cycle or a compliance event, the token savings are irrelevant.

    Use case Pick Why
    Bulk OCR, receipts, invoices DeepSeek V4-Flash-Vision-Exp ~$0.09–$0.17 per 1,000 images; flat 384-token cost per image
    Visual agents, UI automation DeepSeek V4-Flash-Vision-Exp Wins ALE (27.3) and ZeroBench (35.0) head-to-head
    Financial charts, data extraction Claude Opus 4.8 Leads Chartography 65.0 and ApexBench 39.4
    Mixed text-plus-vision reasoning Claude Opus 4.8 DeepSeek’s text performance is unchanged from V4-Flash
    Mid-volume general vision Gemini 3.1 Pro ~$0.50 per 1,000 images with Western data residency
    Production systems needing stability Claude Opus 4.8 or Gemini DeepSeek’s model is explicitly labeled experimental

    What are the catches?

    Three, and none of them are small. The model is experimental, it is closed, and the benchmark margins are inside the noise band. Anyone budgeting a migration on a 1.0-point ZeroBench edge is over-reading the data.

    Start with the “Exp” in the name. DeepSeek ships experimental endpoints and retires them. Building a production pipeline on one is a scheduling risk, not just a quality risk.

    Second, there are no open weights. SiliconANGLE notes DeepSeek “may release a free version later on,” but nothing is confirmed. That kills the self-hosting escape hatch that made Qwen3.8-Max interesting for teams with data-residency constraints.

    Third, the wins are narrow. +1.6 on ALE and +1.0 on ZeroBench are real but slim, and DeepSeek loses ApexBench by a wider 2.9 points than either of its victories. The honest summary is parity, delivered at a fraction of the price — which is still a devastating result for Anthropic’s pricing power.

    Frequently asked questions

    What is DeepSeek V4-Flash-Vision-Exp?

    An experimental multimodal model released on August 21, 2026, extending DeepSeek’s V4-Flash with image understanding. It handles up to 600 images per request with a 1-million-token context window, available only through DeepSeek’s paid API.

    Is the DeepSeek vision model better than Claude Opus 4.8?

    It splits 2-2. DeepSeek wins Agents’ Last Exam (27.3 vs 25.7) and ZeroBench Pass@5 (35.0 vs 34.0). Opus 4.8 wins ApexBench Pass@1 (39.4 vs 36.5) and Chartography (65.0 vs 64.3).

    How much does the DeepSeek vision model cost per image?

    A single image is capped at 384 tokens, costing $0.00008448 off-peak and $0.00016896 at peak. That works out to roughly $0.09 to $0.17 per 1,000 images.

    What are DeepSeek’s peak and off-peak hours?

    Peak hours run 01:00–04:00 and 06:00–10:00 UTC, when input and output rates double to $0.44 and $1.32 per million tokens. Everything outside those windows bills at $0.22 and $0.66.

    Are the weights open?

    No. V4-Flash-Vision-Exp is available only through DeepSeek’s paid developer platform. Reports suggest a free version may follow, but nothing has been confirmed.

    Does it hurt text performance?

    No. DeepSeek’s changelog states the model stays “on par with the official DeepSeek-V4-Flash” on agent reasoning, world knowledge, and other text-based tasks. The vision capability is additive.

    Should I migrate my production vision pipeline?

    Only if you process more than roughly 100,000 images per month and can tolerate an experimental endpoint. Below that volume, the absolute savings do not justify the migration and stability risk.

    The bottom line

    Move your high-volume image workloads to DeepSeek V4-Flash-Vision-Exp and keep your judgment-heavy visual work on Claude Opus 4.8. That is the split the benchmarks actually support, and it is where the money is.

    The specific trigger: if you process more than 100,000 images a month and a wrong answer costs you less than a human review, switch, and schedule the batch off-peak to halve the bill again. Below that volume, or where a visual error creates a compliance event, the $4-per-1,000 Anthropic bill is cheap insurance.

    The larger read is about pricing power, not vision. Anthropic charges $25 per million output tokens on Opus 4.8. DeepSeek charges $0.66 off-peak and just traded benchmark wins with it. The same compression already hit text and coding models, and multimodal was the last premium tier standing.

    Sources

  • Nvidia AVO Hits 100% on ARC-AGI-3. The Model Alone Scored 30.2%.

    Nvidia AVO — Agentic Variation Operators — scored 100.00 RHAE on the ARC-AGI-3 public set on August 21, 2026, clearing all 183 levels across 25 environments in 6,624 actions, roughly 12% fewer than the VISTA baseline. The same base model, Claude Opus 5, scores 30.2% on its own. The harness did the work, and that changes where agent money goes.

    What is Nvidia AVO?

    Nvidia AVO stands for Agentic Variation Operators. It is not a model. It is a general-purpose coding-agent system that wraps an existing frontier model in a loop — inspect, plan, implement, evaluate — plus persistent memory and a supervisor that intervenes when progress stalls. Nvidia published the results on August 21, 2026.

    The base model inside the winning run was Anthropic’s Claude Opus 5. Nvidia also ran limited experiments with GPT-5.6 Sol on a subset of games, and labeled those findings preliminary.

    That detail is the whole story. Nvidia did not train a better reasoner. It built better scaffolding around someone else’s reasoner.

    How the AVO loop works

    AVO runs a four-step cycle: inspect the current context, plan a change, implement it, then evaluate the result against the environment.

    Two additions separate it from a standard agent loop, according to Nvidia’s technical blog:

    • Persistent memory that carries forward prior implementations, evaluation results and reasoning across the whole run, not just the current context window.
    • A supervision mechanism that watches the trajectory and redirects the agent when it detects the run has stopped making progress.
    • Variation operators that generate structured alternatives rather than retrying the same failed approach.

    Why the supervisor is the expensive part

    Long-horizon agent failure is rarely a single wrong answer. It is a slow drift — the agent loops on a dead approach and burns tokens without noticing.

    A supervisor that detects stagnation is cheap to describe and hard to build. It is also the component least likely to transfer cleanly to another benchmark.

    What is ARC-AGI-3 and why does a 100% score matter?

    ARC-AGI-3 is ARC Prize’s interactive reasoning benchmark: 25 pixel-art puzzle environments containing 183 public levels. Agents get no instructions, no rules and no goal labels. They must infer the mechanics purely by playing. When the benchmark launched, humans cleared 100% of environments and the best AI managed 0.37%.

    That 0.37% figure is why this result registered. ARC-AGI-3 was designed as the benchmark models could not touch.

    How RHAE scoring works

    The metric is RHAE — Relative Human Action Efficiency. It combines task completion with how many actions the agent needed per level, measured against initial human performance, then aggregates across every level and environment.

    So a 100.00 does not just mean “finished everything.” It means finishing everything at roughly human action efficiency. ARC Prize published its human performance dataset specifically so this number would have a floor to sit on.

    How much did the harness add versus the raw model?

    The gap is 30.2% to 100.00 — the same model class, wrapped differently. ARC Prize reported Claude Opus 5 at 30.2% on ARC-AGI-3 in July 2026, which it called a genuine reasoning leap at the time. Nvidia’s harness took that model to a clean sweep of the public set.

    SystemARC-AGI-3 resultActions usedReported by
    Best AI at benchmark launch0.37%—ARC Prize
    Claude Opus 5 (bare model)30.2%—ARC Prize, July 2026
    VISTA baseline agentCleared same level sets7,542Nvidia
    Nvidia AVO (Claude Opus 5 inside)100.00 RHAE, all 183 levels6,624Nvidia, Aug 21 2026

    For context on the base model’s ceiling elsewhere: Claude Opus 5 at maximum reasoning effort scores 97.5% on ARC-AGI-1 and 90.4% on ARC-AGI-2, per The New Stack. ARC-AGI-3 was the one that stayed hard.

    Nvidia’s own framing, from the blog post: “The model matters, but the model is not the entire agent.”

    Is the 100% score real, or is it benchmark theater?

    It is real, and it is narrower than the headline suggests. The score covers the ARC-AGI-3 public set only — not the semi-private or private competition sets that ARC Prize maintains precisely to catch overfitting. Nvidia says so in its own post.

    The public set is not the hidden exam

    Benchmark designers split datasets for a reason. A public set is a practice test with the answers eventually leaking into the ecosystem around it.

    One commenter on Nvidia’s announcement put it bluntly, as flagged in explainX’s write-up: “I would not file this as solved AGI… if you post 100 like it is the hidden exam.” Until AVO posts a semi-private number, that objection stands.

    These are not controlled ablations

    Nvidia explicitly labels its comparisons as not controlled ablations. That matters more than it sounds.

    The AVO-versus-VISTA action count — 6,624 against 7,542 — varies agent backends, observation formats, memory systems and reasoning settings all at once. The 100.00-versus-30.2% comparison swaps the entire system architecture and the reasoning-effort setting simultaneously.

    Neither number isolates how much the harness itself contributed. The honest reading is “a well-built harness closed a very large gap,” not “the harness is worth exactly 70 points.”

    Nvidia also disclosed no compute cost, no token usage and no wall-clock runtime for the 6,624 actions. For anyone pricing an agent product, that is the number that actually matters — and it is missing.

    Has AVO done anything useful outside a puzzle benchmark?

    Yes, and this is the part investors should read twice. Nvidia ran AVO continuously for seven days on GPU-kernel optimization, exploring more than 500 optimization directions. The system produced kernels that beat FlashAttention-4 by up to 10.5% on NVIDIA DGX B200 hardware.

    FlashAttention is not a soft target. It is hand-tuned infrastructure that the entire industry’s inference economics rest on.

    A 10.5% kernel improvement compounds across every token served on that hardware. If it holds in production, it is worth more to Nvidia than the benchmark headline — and it lands in the same week the company has been buying capability outright elsewhere.

    Who wins and who loses financially?

    The winner is whoever owns the orchestration layer. If a 30% model becomes a 100% agent through harness design, then value is accruing above the weights, not inside them. That is bad news for anyone whose entire moat is a checkpoint.

    Nvidia is climbing the stack

    AVO did not appear in isolation. On the same day, Nvidia paid $6 billion to license Poolside’s model-development software and invested $1 billion more in the startup, according to PYMNTS.

    A chip company publishing frontier agent architecture and licensing a model factory in the same 24 hours is not a coincidence. It is a company that has watched its customers capture the margin its silicon creates — the same dynamic behind its recalculated OpenAI data center guarantee.

    Model labs keep pricing power, for now

    Note who supplied the brain: Anthropic. AVO’s best run needed Claude Opus 5, and the harness could not manufacture reasoning that was not already there — the 0.37% launch-day figure is proof that scaffolding alone does nothing on a weak model.

    So frontier labs still sell the scarce input. What they lose is the claim that the model is the product, which shows up quickly in cheaper models closing capability gaps.

    Agent startups just got a harder question

    Three practical implications for anyone building or funding an agent company:

    1. Harness design has not hit diminishing returns. A 30-to-100 jump says the scaffolding layer is still under-engineered — which is opportunity and commoditization risk in the same sentence.
    2. Your differentiator may be a blog post away from replication. Persistent memory plus a stagnation supervisor is a describable architecture, not a trade secret.
    3. Nvidia is now a potential competitor, not just a supplier. It has the hardware, the capital, and as of August 21, published frontier agent research.

    The cost question decides all three. Running a supervised, memory-heavy loop for 6,624 actions is not free, and the economics look very different depending on whether the underlying tokens cost $2 or $60 per million — the same math that drives coding-agent unit costs and inference vendor selection.

    Frequently asked questions about Nvidia AVO

    Is Nvidia AVO a new AI model?

    No. AVO is an agent system — a harness — that runs on top of existing frontier models. The reported 100.00 RHAE run used Claude Opus 5 as its base model.

    Did Nvidia AVO solve AGI?

    No. The score covers ARC-AGI-3’s public set of 183 levels across 25 environments. ARC Prize also maintains semi-private and private sets, and AVO has not posted a result on those.

    What does RHAE mean?

    Relative Human Action Efficiency. It scores both whether an agent completes a level and how many actions it needed relative to initial human performance, aggregated across the benchmark.

    Can developers use AVO today?

    Nvidia’s August 21 post describes the architecture and results. It does not announce a code or weights release, so treat AVO as published research rather than a shippable dependency.

    How much does an AVO run cost?

    Nvidia did not disclose compute cost, token usage or wall-clock time for the benchmark run. Without those figures, the result cannot be compared on a cost-per-task basis against cheaper agent harnesses.

    What was the FlashAttention-4 result?

    Running for seven days across 500-plus optimization directions, AVO produced GPU kernels that outperformed FlashAttention-4 by up to 10.5% on NVIDIA DGX B200 hardware.

    Does this make Claude Opus 5 look better or worse?

    Both. The model was capable enough to be driven to 100.00 RHAE, and weak enough on its own to score 30.2%. The delta belongs to the harness, not the checkpoint.

    The bottom line

    Nvidia AVO is the most important agent result of the month, and the headline number is the least interesting part of it.

    A 100.00 on a public set with no controlled ablations and no disclosed cost is a demonstration, not a benchmark victory. Anyone treating it as “ARC-AGI-3 is solved” is reading a press release as a result.

    What survives scrutiny is the gap: 30.2% to 100.00, same model, different scaffolding. That gap is the clearest evidence yet that in 2026 the agent layer, not the model layer, is where the remaining engineering leverage sits.

    And the FlashAttention-4 kernels are the tell. Nvidia did not build AVO to win a puzzle leaderboard. It built AVO to make its own hardware faster — and, at a moment when record sums are being raised to finance AI chip capacity, to stop being only the company that sells the machines.

    Sources

  • Nvidia Poolside Deal: $6 Billion for a Model Factory and 109 Staff

    Nvidia is paying Poolside $6 billion to license its model-building software and hiring 109 of the startup’s staff, according to Newcomer, which broke the story on August 20, 2026. A separate $1 billion investment values what remains at $12 billion pre-money. Nvidia shares closed the week down roughly 5%. No company legally changes hands.

    The Nvidia Poolside deal is the third time in twelve months that the world’s most valuable chipmaker has bought a startup without buying a startup. It is becoming a template.

    What exactly is the Nvidia Poolside deal?

    Three transactions in one package. Nvidia pays $6 billion for a non-exclusive license to Poolside’s “model factory,” extends offers to 109 employees, and invests $1 billion at a $12 billion pre-money valuation. Poolside keeps its name, its three founders, and its corporate independence.

    The “model factory” is not a model. It is the system Poolside built to produce models — the training pipeline, the data infrastructure, the orchestration layer.

    Nvidia is buying the assembly line, not the car.

    Bloomberg confirmed the terms on August 20, citing Newcomer’s reporting. The Information reported the same package the following day.

    The deal terms, line by line

    Component Terms Source
    Technology license $6 billion, non-exclusive Newcomer, Aug 20, 2026
    Equity investment $1 billion Bloomberg
    Valuation of remaining entity $12 billion pre-money Newcomer
    Staff receiving Nvidia offers 109 employees The Next Web
    Founders staying with Poolside 3 The Next Web
    Proceeds distributed to investors By end of 2027 Poolside investor letter
    Prior Nvidia commitment to Poolside Up to $1 billion (October 2025) The Next Web

    Why did Poolside sell its model factory?

    Because it could not afford the chips. Poolside’s investor letter, quoted by The Next Web, describes a financing failure with a hard deadline: the company needed $2 billion in six weeks to pay for a 40,000-GPU cluster, missed the window, and lost the allocation.

    The letter is unusually blunt. “We had a 6 week window in which to raise $2 billion dollars to pay for a 40,000 GB300 cluster coming online in January,” it reads. “We didn’t close it in time, and we lost the cluster.”

    Poolside’s own assessment: a frontier-competitive model needs 10,000 to 20,000 of those chips today, and materially more next year.

    That is the whole story of the independent model lab in 2026, compressed into two sentences. The research talent is not the constraint. The capital stack is.

    • The gap: $2 billion needed in six weeks, against a $12 billion pre-money valuation
    • The consequence: allocation forfeited, frontier ambitions shelved
    • The pivot: Poolside moved from coding agents into data center operations and open-weight releases before the deal
    • The buyer: the company that sells the chips it could not pay for

    CEO Eiso Kant and two co-founders remain. The people who actually built the thing — fewer than 70 on the model itself, under 115 across engineering and research combined — largely go to Nvidia.

    How does this compare to Nvidia’s Groq and Enfabrica deals?

    It is the same structure at a different price. Across three transactions, Nvidia has committed roughly $27 billion to license technology and absorb teams while leaving the original corporate entities standing. Groq was the largest at about $20 billion. Enfabrica was roughly $900 million.

    Target Reported value What Nvidia received
    Enfabrica ~$900 million License plus networking team
    Groq ~$20 billion Non-exclusive design license, founders, most staff
    Poolside $6B license + $1B equity Model-factory license, 109 staff
    Combined ~$27 billion Three teams, zero acquisitions

    The pattern is deliberate enough that it now has a name in the trade press: the reverse acquihire. Buy the license, hire the people, leave the shell.

    Is the reverse acquihire an antitrust workaround?

    Two US senators have already said so in writing. On March 23, 2026, Elizabeth Warren and Richard Blumenthal wrote to Jensen Huang about the Groq deal, arguing that Nvidia “has effectively acquired Groq in all but name” by licensing its technology and hiring its key employees.

    The letter cites Nvidia’s roughly 90% share of the GPU market and warns the structure “could stifle competition, further entrenching NVIDIA’s dominance in the AI chip industry.”

    The senators’ core objection is procedural. A conventional acquisition triggers premerger notification and agency review. A license plus a hiring spree does not — even when the economic result is indistinguishable.

    The FTC and DOJ retain authority to investigate consummated transactions regardless of filing status. Whether they will is a different question. As of this writing, no public enforcement action has been announced against any of the three deals.

    Read the full Warren-Blumenthal letter for the argument in the senators’ own words.

    What are the skeptical questions about the $6 billion price?

    Start with the arithmetic. Nvidia is paying $6 billion for a non-exclusive license to software built by fewer than 115 people at a company that just failed to raise $2 billion. That is roughly $55 million per engineer hired, and the license does not stop Poolside from licensing the same technology elsewhere.

    Non-exclusive is the word doing the most work in this deal.

    Second question: what is Nvidia actually short of? It is not model-training expertise — Nvidia has built plenty. The likelier answer is speed. Buying a working pipeline compresses years into a quarter.

    Third: the money moves in a familiar circle. Nvidia committed up to $1 billion to Poolside in October 2025. Poolside spent on Nvidia hardware. Nvidia now pays $6 billion back, some of which flows to investors by end of 2027, and takes another $1 billion equity position. Revenue and investment are increasingly hard to separate on this balance sheet.

    We flagged the same circularity concern when Nvidia cut its OpenAI data center guarantee from $250 billion to $120 billion — a revision that suggested even Nvidia has limits on how much demand it will underwrite itself.

    The market noticed. Nvidia shares fell about 5% over the week of the announcement, closing Friday down 0.9%, though the stock remains up 14.5% year to date.

    Why this matters for the AI market

    The Nvidia Poolside deal marks the point where compute access stopped being a competitive advantage and became a gate. Poolside had the talent, the models, and a $12 billion valuation. It still could not clear a $2 billion payment on schedule, and that alone ended its frontier ambitions.

    For investors, three implications follow.

    1. The exit landscape has changed. A reverse acquihire returns capital without an acquisition premium, an IPO, or regulatory review. Cap tables should price that in.
    2. Valuation and viability have decoupled. A $12 billion paper valuation did not translate into $2 billion of callable cash in six weeks.
    3. Nvidia is consolidating the stack quietly. Roughly $27 billion across three deals, none of which required a merger filing.

    Compare this to the conventional route: Stripe paid an estimated $7 billion and actually bought the company when it acquired OpenRouter. Nvidia is getting comparable strategic value for less, with less scrutiny.

    Meanwhile the debt markets are doing their own version of the same trade — Broadcom is arranging up to $100 billion to finance AI chip infrastructure. Capital is chasing compute from every direction at once.

    This post is reporting and analysis, not financial advice.

    Frequently asked questions about the Nvidia Poolside deal

    Short answers to the questions readers are asking about the structure, the price, and what happens next.

    Did Nvidia acquire Poolside?

    No. Poolside remains an independent company with its three founders and its own board. Nvidia licensed technology and hired staff. No change of control occurred.

    How much is Nvidia paying in total?

    $6 billion for the non-exclusive license plus a $1 billion equity investment — $7 billion combined, per Newcomer and Bloomberg reporting from August 20, 2026.

    How many Poolside employees are joining Nvidia?

    109 received offers, according to The Next Web. Poolside had fewer than 115 people across engineering and research in total.

    What is a “model factory”?

    The infrastructure and process Poolside built to train AI models — pipelines, data systems, orchestration. Nvidia licensed the production system rather than any individual model.

    Why is this called a reverse acquihire?

    A normal acquihire buys a company to get its people. Here Nvidia gets the people and the technology while the company survives, avoiding merger review.

    Has this structure faced regulatory pushback?

    Senators Warren and Blumenthal challenged Nvidia’s similar $20 billion Groq deal in a March 23, 2026 letter, calling it an acquisition “in all but name.” No enforcement action has followed publicly.

    What happens to Poolside now?

    It continues with $1 billion in fresh capital at a $12 billion pre-money valuation and plans to distribute the $6 billion license proceeds to investors by end of 2027.

    The bottom line

    Nvidia has found a way to buy companies that does not look like buying companies, and it has now used it three times for roughly $27 billion. The Poolside deal is the cleanest example yet: a startup that could not fund its own chips sold the machine that would have used them, to the company that makes them.

    Expect two things next. More labs will take this exit — the economics of independent frontier training are brutal, and a license-plus-hire returns capital fast. And expect the structure to draw a formal response from Washington, because three deals is a pattern, not a coincidence.

    Watch for whether the FTC opens a review. That is the variable that decides whether this template survives 2027.

    Sources

  • Broadcom AI Debt Deal: Up to $100 Billion to Fund Anthropic Chips

    Broadcom is seeking as much as $100 billion in debt to finance custom AI chips for Anthropic and other labs, according to Bloomberg. The Broadcom AI debt deal would layer a senior secured tranche of $60–70 billion, partly guaranteed by Broadcom, on top of roughly $30 billion of junior debt. CNBC puts the likely total nearer $70–80 billion. Blackstone and Apollo are the lenders.

    What is the Broadcom AI debt deal?

    It is a private credit financing, not a stock sale. Broadcom is arranging debt that funds AI infrastructure for its own customers — with Anthropic named as the primary beneficiary. Bloomberg reported the talks on August 20, 2026, citing people with knowledge of the matter. Broadcom declined to comment.

    The structure matters more than the headline. Broadcom does not simply sell chips here. It helps assemble the capital that lets a customer buy them.

    That is vendor financing, and it is now the dominant pattern in AI infrastructure.

    How the tranches are structured

    Reporting differs on size, which is itself informative — the deal is not closed. Bloomberg and CNBC describe two different splits.

    Component Bloomberg (Aug 20) CNBC (Aug 21)
    Senior secured tranche $60–70 billion ~$45 billion
    Junior tranche ~$30 billion ~$35 billion
    Indicated total Up to $100 billion $70–80 billion
    Broadcom guarantee Portion of senior debt Not specified
    Named lenders Blackstone, Apollo Blackstone, Apollo

    A $30 billion spread between two credible outlets on the same deal, one day apart, is a reminder that these numbers are being shopped, not signed.

    How much has Broadcom already raised for this platform?

    $35 billion. In June 2026, Broadcom, Apollo and Blackstone launched the AI XPV Platform, with Apollo leading a $35 billion capital solution and Blackstone’s credit and insurance business as anchor investor. The new raise is an expansion of that vehicle, not a fresh idea.

    The June announcement set the ambition: more than 20 gigawatts of AI deployments through 2028, with over 1 gigawatt of initial capacity earmarked for Anthropic.

    Twenty gigawatts is roughly the output of twenty nuclear plants.

    “This strategic Platform with Apollo and Blackstone synchronizes the world’s most sophisticated capital with Broadcom’s advanced technological roadmap,” Broadcom CEO Hock Tan said in the June 9 release.

    Apollo President Jim Zelter framed it as a bet on the customer as much as the supplier: the investment “reflects our conviction in Broadcom’s technology leadership and Anthropic’s frontier roadmap.”

    Why is Anthropic at the center of the deal?

    Because Anthropic has become Broadcom’s largest custom-silicon commitment. Broadcom expects AI chip revenue above $100 billion next year, and Anthropic is projected to account for more than 40% of it, per SiliconANGLE’s reporting on the financing.

    Anthropic’s own numbers explain the appetite. Its annualized revenue run rate passed $65 billion in August 2026, Axios and Bloomberg reported, ahead of a widely expected IPO.

    The compute schedule is aggressive:

    • 1 gigawatt of capacity delivered in 2026
    • 3 gigawatts planned for 2027
    • Initial deployments at Fluidstack-operated sites from mid-2026
    • OpenAI’s first Broadcom-built custom chip targeted for 2027
    • Meta’s MTIA accelerators already shipping

    Broadcom’s role is to design the XPUs and networking that let labs escape Nvidia’s pricing. The debt is what makes that escape affordable before the revenue arrives.

    How does this compare to Nvidia’s OpenAI financing?

    Closely — which is the point. Three days before the Broadcom news, Nvidia agreed to back OpenAI’s Ohio data center with up to $105 billion, according to Bloomberg and UPI. Two chip suppliers, two customer-financing packages, one week.

    Metric Broadcom / Anthropic Nvidia / OpenAI
    Reported size Up to $100 billion (sought) Up to $105 billion (agreed)
    Date reported Aug 20–21, 2026 Aug 17, 2026
    Form Senior + junior debt, partial guarantee Financing backstop / guarantee
    Capital partners Blackstone, Apollo SB Energy, SoftBank
    Capacity 20+ GW through 2028 (platform) 4.25 GW initial, 3.75 GW option
    Online From 2026 2028

    We covered the Nvidia side when the company cut its OpenAI data center guarantee from $250 billion to $120 billion. The direction of travel since then has been more customer financing, not less.

    Who profits from the AI debt deal?

    Private credit does, first and most reliably. Blackstone and Apollo earn contracted yield on infrastructure debt that is secured against chips and computing capacity, and they get paid whether or not Anthropic’s models win.

    Broadcom profits second. It converts a customer’s capital constraint into a booked order, and it does so without spending its own balance sheet — except for the guarantee.

    That guarantee is the part investors should read twice.

    The circularity problem

    If Broadcom guarantees a portion of the senior tranche, it is underwriting demand for its own product. Revenue recognized today rests partly on a liability Broadcom would owe tomorrow if the customer stumbles.

    This is not fraud and it is not new — telecom vendors did it in the late 1990s. It is simply a structure that looks excellent while growth holds and ugly the moment it does not.

    Broadcom shares rose slightly more than 1% on Friday, August 21, per CNBC. The market is not pricing much risk into this.

    Why this matters

    AI capital formation has moved from venture equity to leveraged infrastructure. That is a different asset class with different failure modes.

    Equity investors lose money slowly and quietly. Debt has covenants, maturities and forced sales.

    Three implications for anyone tracking the AI trade:

    1. Chip demand is now credit-dependent. A tightening in private credit spreads would hit AI capex faster than any drop in model quality.
    2. Broadcom is becoming a financing company with a fabless chip business attached. Its risk profile is drifting away from pure semiconductors.
    3. Anthropic’s IPO math gets more complex. Compute secured through supplier-arranged debt is cheaper up front and heavier later.

    For context on the customer’s valuation, see our piece on Anthropic’s $2 trillion mark and its $6 billion Decart deal, and on where inference dollars are actually landing, our Cerebras vs Groq cost comparison. It also rhymes with the private-market repricing we saw in Databricks’ $190 billion round.

    This post is reporting and analysis, not financial advice.

    Frequently asked questions

    How much is Broadcom raising?

    Between $70 billion and $100 billion, depending on the report. CNBC says $70–80 billion; Bloomberg says more than $60 billion with a total that could reach $100 billion. Nothing is finalized.

    Is Broadcom borrowing this money itself?

    No. Broadcom is arranging the financing and may guarantee part of the senior tranche. The debt is raised through the AI XPV platform with Apollo and Blackstone.

    Who are the lenders?

    Blackstone and Apollo Global Management, the same two firms that led the $35 billion tranche announced in June 2026.

    What does Anthropic get?

    Access to Broadcom custom XPUs and networking, plus the data center capacity to run them. Initial deployments target more than 1 gigawatt, rising to a reported 3 gigawatts in 2027.

    How big is Broadcom’s AI business?

    Broadcom reported $8.4 billion in AI revenue in its fiscal first quarter of 2026 on $19.31 billion total, and guided to roughly $10.7 billion in AI chip revenue the following quarter. Hock Tan has said the company has “line of sight” to more than $100 billion in AI chip revenue in 2027.

    Does this threaten Nvidia?

    At the margin. Custom silicon is how large labs reduce Nvidia dependence. But Nvidia is running the same playbook, backing OpenAI’s Ohio campus with up to $105 billion.

    What is the main risk?

    Concentration. One customer is projected to drive more than 40% of Broadcom’s AI chip revenue while also being the borrower whose debt Broadcom partly guarantees.

    The bottom line

    The Broadcom AI debt deal is the clearest sign yet that the AI buildout has outgrown equity. When two suppliers arrange roughly $200 billion of customer financing in a single week, the constraint is no longer conviction — it is balance sheet.

    Watch three things next: whether the senior tranche prices near the reported $60–70 billion or closer to CNBC’s $45 billion, how much of it Broadcom guarantees, and whether Anthropic’s IPO filing discloses the obligations attached to this capacity.

    The chips are the easy part now. The financing is the story.

    Sources

  • Cerebras vs Groq: Which Fast Inference API Is Worth the Money

    Cerebras vs Groq comes down to one trade. Cerebras serves GPT-OSS-120B at 1,641 tokens per second for $0.75 per million output tokens. Groq serves the same model at roughly 500 for $0.60. You pay about 25% more on output for roughly three times the speed. Buy Cerebras when a human or an agent is waiting. Buy Groq for batch work and overnight jobs.

    That trade just got sharper. On August 18, 2026, Cerebras announced the CS-4, a system it claims runs GPT-OSS-120B at more than 4,400 tokens per second per user.

    If that number survives contact with production traffic, the speed gap stops being a nice-to-have and starts being a product feature you can charge for.

    What changed in the Cerebras vs Groq race this week?

    Cerebras shipped a new generation of silicon and Nvidia now owns its main rival’s technology. Those two facts reshape the fast-inference market. The CS-4 raises Cerebras’ ceiling; the Nvidia-Groq deal means Groq’s aggressive pricing is no longer set by a scrappy independent.

    The CS-4 numbers that matter

    Per the Cerebras announcement, the CS-4 delivers 750 PFLOPS of AI compute against the CS-3’s 125 PFLOPS. Memory bandwidth jumps from 21.6 to 129.6 petabytes per second.

    The WSE-3 Turbo processor behind it packs 4 trillion transistors and 900,000 AI cores across 46,225 square millimeters of silicon, with 44GB of on-chip SRAM.

    Wafer-to-wafer latency drops from 5 microseconds to 2. Cerebras also claims up to 10x the throughput per watt versus the CS-3, and support for models above 50 trillion parameters.

    The CS-4 product page adds a second claim worth watching: more than 1,000 tokens per second on models exceeding 10 trillion parameters. First shipments began in Q3 2026.

    CTO Sean Lie framed the pitch in agent terms: “Being 30 times faster gives an agentic system room for significantly more reasoning, verification, or tool use in the same wall-clock time.”

    Why Nvidia now sits on both sides

    Groq is no longer an independent challenger. CNBC reported on December 24, 2025 that Nvidia agreed to buy Groq’s assets for about $20 billion — its largest deal on record.

    The Groq API still runs and still undercuts Cerebras. But the pricing that made Groq attractive is now a line item inside the company that also sells the GPUs Groq was built to beat.

    That matters for anyone building a business on a specific cost per token. Cerebras is the last large pure-play fast-inference vendor with its own silicon and its own incentive to keep prices down.

    How much does fast AI inference cost per million tokens?

    Cerebras is the most expensive way to run GPT-OSS-120B among mainstream providers. Groq sits mid-pack. Commodity GPU serverless tiers cost a fraction of both. The spread on the identical open-weights model is roughly 7x on output tokens, which is far wider than most teams assume.

    Here is the pricing snapshot for GPT-OSS-120B as tracked by PricePerToken on August 21, 2026, with measured speed where it is published.

    Provider Input $/M Output $/M 1M in + 1M out Measured output speed
    Cerebras $0.35 $0.75 $1.10 1,641 tok/s
    Groq $0.15 $0.60 $0.75 ~500 tok/s
    SambaNova $0.14 $0.95 $1.09 Not published
    Together AI $0.15 $0.60 $0.75 Not published
    Amazon Bedrock $0.15 $0.60 $0.75 Not published
    Baseten $0.10 $0.50 $0.60 Not published
    Google $0.09 $0.36 $0.45 Not published
    Fireworks $0.10 $0.10 $0.20 Not published
    DeepInfra $0.037 $0.170 $0.207 Not published
    OpenRouter $0.030 $0.170 $0.200 Not published
    Cerebras speed from Artificial Analysis; Groq speed from CloudZero, May 2026.

    Run a balanced million-in, million-out workload and Cerebras costs $1.10 against Groq’s $0.75 — a 47% premium. On output tokens alone the gap narrows to 25%.

    Both are expensive next to Fireworks at $0.20 or the OpenRouter route at $0.20 for the same weights.

    The discount lever most teams forget

    Groq’s list price is not its real price. CloudZero notes that Groq’s Batch API and prompt caching each cut rates by 50%, and the two stack to roughly 25% of on-demand pricing.

    Applied to GPT-OSS-120B, that pushes effective output cost toward $0.15 per million. Cerebras publishes no equivalent public stacking discount; its pricing page lists a $5 free trial, a $10 self-serve developer tier, and custom enterprise rates.

    For any workload that tolerates a batch window, Groq is not 20% cheaper. It is closer to 5x cheaper.

    Which is faster in practice, Cerebras or Groq?

    Cerebras wins today, and not narrowly. Independent measurement puts Cerebras at 1,641 tokens per second on GPT-OSS-120B with 0.46 seconds to first token. Groq’s published figure for the same model is around 500 tokens per second. That is a 3.3x throughput advantage before the CS-4 ships at volume.

    Artificial Analysis also clocks Cerebras at 1,402 tokens per second on Gemma 4 31B in reasoning mode, at a blended $0.24 per million.

    What the speed gap costs in real money

    Take a 100,000-token generation — a long agent trace or a full document rewrite.

    • Cerebras today: 61 seconds at 1,641 tok/s, costing $0.075 in output tokens.
    • Groq today: 200 seconds at 500 tok/s, costing $0.060.
    • Cerebras CS-4 claim: 23 seconds at 4,400 tok/s.
    • The math: 1.5 cents buys back 139 seconds — roughly 11 cents per minute of latency removed.

    Eleven cents a minute is trivial if a customer is watching a cursor blink. It is indefensible if the job runs at 3 a.m. and nobody reads the output until morning.

    One honest caveat: the 4,400 tok/s figure is a vendor claim tied to hardware that only started shipping this quarter. The 1,641 figure is measured on the live API. Do not budget against the former.

    Is Cerebras worth the premium for AI agents?

    Yes, for interactive and multi-step agent work — and this is the only case where the premium clearly pays. Agent loops multiply latency: ten sequential tool calls at 200 seconds each is a 33-minute task. The same loop at Cerebras speed finishes in about 10 minutes.

    That compounding is the whole argument. A single completion at 500 tokens per second feels fine. Twenty of them chained behind a task does not.

    Who should buy what

    Use case Pick Why
    Live chat, copilots, voice Cerebras 0.46s TTFT and 1,641 tok/s; latency is the product
    Multi-step agents with tool calls Cerebras Per-step latency compounds across the loop
    Overnight batch, evals, data labeling Groq (Batch API) Stacked discounts reach ~25% of list
    High-volume, cost-capped production Fireworks / DeepInfra $0.20 per 1M in + 1M out on identical weights
    Closed frontier models Neither Both serve open weights only
    Sustained 24/7 single-model load Self-host Fixed GPU cost beats per-token above a break-even

    That last row matters more as open weights close the quality gap. We ran the self-hosting break-even in our Qwen3.8-Max open weights breakdown, and the logic holds here.

    Note the hard limit on both vendors: neither serves closed frontier models. If your stack depends on the newest proprietary coder, this comparison does not apply — see our GLM-5.3 vs DeepSeek V4 Pro comparison for the open-weight options that do run here.

    What do the financials say about who wins?

    Cerebras has the better technology and the shakier income statement. It went public on May 14, 2026, raising $5.5 billion. The stock priced at $185, opened at $385 for a 108% pop, and closed at $311 — a $66 billion valuation. Three months later the market repriced it hard.

    The Q2 miss

    On August 12, 2026, Cerebras reported Q2 revenue of $180.1 million against a $193.6 million consensus, per Investing.com. Adjusted EPS came in at a $2.98 loss versus an expected $0.18 loss. Shares fell 14% after hours.

    Core revenue still grew 103% year over year to $209.9 million, and the company guided FY2026 core revenue to $880–890 million. Growth is not the problem. Margin is: guided core operating margin sits at negative 19% to negative 17%.

    Why that should affect your buying decision

    A vendor losing money on every wafer has two exits: raise prices or get acquired. Cerebras’ 2025 revenue was $510 million on 76% growth with $237.8 million of net income, so the balance sheet is not fragile — but the 2026 trajectory is being funded, not earned.

    TrendForce values the company’s three-year OpenAI partnership at over $20 billion. That is concentration risk dressed as a moat.

    The broader point TrendForce makes is the one to internalize: inference is a recurring cost tied directly to revenue, while training is a one-time R&D expense. Its example is brutal — Taalas’ HC1 delivers Llama 3.1 8B at 0.75 cents per million tokens against 3.79 cents on an Nvidia B200.

    Specialized silicon is roughly five times cheaper per token than general-purpose GPUs at that scale. That is why the price you lock in today is unlikely to be the price in twelve months.

    Frequently asked questions about Cerebras vs Groq

    Is Cerebras faster than Groq?

    Yes. Artificial Analysis measures Cerebras at 1,641 tokens per second on GPT-OSS-120B; Groq’s published figure for the same model is around 500. Cerebras also leads on time to first token at 0.46 seconds.

    Is Groq cheaper than Cerebras?

    Yes. On GPT-OSS-120B, Groq charges $0.15 input and $0.60 output per million tokens versus Cerebras at $0.35 and $0.75. With Groq’s stacked batch and cache discounts, the effective gap widens sharply.

    Does Nvidia own Groq now?

    Nvidia agreed in December 2025 to acquire Groq’s assets for about $20 billion, CNBC reported. The Groq API continues to operate and continues to publish its own pricing.

    Can I run Claude or GPT-5 on Cerebras or Groq?

    No. Both providers serve open-weights models only — GPT-OSS, Llama, Qwen, Gemma and similar. Closed frontier models stay on their vendors’ own APIs.

    When does the CS-4 speed actually arrive for API users?

    Cerebras says first CS-4 shipments began in Q3 2026. The 4,400 tokens per second per user figure is a vendor claim on new hardware, not yet an independently measured API result.

    Is paying for faster inference ever worth it?

    Only when latency is visible to a customer or compounds across an agent loop. At roughly 11 cents per minute of latency removed, speed is cheap for interactive products and pure waste for background jobs. We covered the same trade at the frontier-model layer in OpenAI Ultrafast vs Claude Fast Mode.

    The bottom line

    Route by whether something is waiting. If a human or an agent loop blocks on the token stream, Cerebras is worth its 25% output premium — 3.3x measured throughput for that price is one of the better deals in AI infrastructure, and the CS-4 should widen it.

    If nothing is waiting, Cerebras is a rounding-error upgrade you are overpaying for. Send batch and evaluation traffic to Groq’s Batch API at roughly a quarter of list, or to Fireworks and DeepInfra at $0.20 per million in and out.

    The strategic read is less comfortable. Nvidia owns Groq’s technology, Cerebras is losing money at negative 17% to 19% core operating margin, and specialized silicon is already showing five-fold cost advantages per token. Sign nothing longer than twelve months.

    Sources

  • Unitree IPO Pops 629%: China’s Robot Maker Hits $66 Billion

    Unitree Robotics opened 629% above its IPO price on Shanghai’s STAR Market on August 19, 2026, briefly valuing the humanoid robot maker at about 445 billion yuan ($66 billion). It closed up 460% at 845 yuan. The company raised 6.1 billion yuan ($904 million) on 2025 revenue of just 1.7 billion yuan — roughly 210 times sales, per Forbes.

    The Unitree IPO is the loudest thing that has happened in robotics financing this year, and the numbers behind it are stranger than the headline pop suggests.

    China now has a listed humanoid robot maker worth more than Baidu. It sells fewer than 20,000 robots in total. Both of those statements are true at the same time.

    What happened in the Unitree IPO?

    Unitree Robotics listed on the Shanghai Stock Exchange’s STAR Market on August 19, 2026, priced at 150.80 yuan per share. The stock opened at 1,100 yuan — a 629% gain — then gave back most of the spike to close at 845 yuan, up 460%, according to the South China Morning Post.

    It is the first pure-play humanoid robot maker to list anywhere. That scarcity is doing a lot of work in the price.

    The float was small by design. Unitree sold 40.45 million shares, about 10% of its enlarged capital, raising 6.1 billion yuan — roughly $904 million, as Bloomberg reported ahead of the debut.

    The debut in numbers

    Metric Figure Source
    IPO price 150.80 yuan/share SCMP / The Standard
    Opening price 1,100 yuan (+629%) SCMP
    Closing price 845 yuan (+460%) SCMP / Bloomberg
    Shares sold 40.45 million (~10% of capital) Forbes
    Amount raised 6.1 billion yuan (~$904M) Bloomberg
    Market cap at open ~444.9 billion yuan (~$66B) The Standard / Fortune
    Market cap at close ~342 billion yuan (~$48B) SCMP
    Retail oversubscription More than 5,500x The Standard
    First-day turnover 23.2 billion yuan SCMP

    One detail is worth pausing on. Even the professional coverage could not agree on where the stock finished: CNBC reported a 542% close and Forbes a 487% close, while SCMP, Bloomberg, Quartz and Fortune all landed on 460%. The arithmetic favors 460% — 845 divided by 150.80 is a 5.6x return. When a debut moves this fast, the tape itself becomes hard to read.

    Why did the Unitree IPO open 629% higher?

    Because supply was engineered to be tiny and demand was not. The retail tranche was oversubscribed more than 5,500 times, and SCMP counted roughly 9.8 million retail accounts chasing about 9.7 million available shares. That is close to one share per applicant.

    China’s IPO lottery system converts that imbalance directly into a first-day gap. The Standard calculated that a single 500-share allocation was worth about 474,600 yuan in paper profit at the open.

    The list of people who got in at 150.80 yuan was, by construction, very short. Everyone else had to buy from them.

    This is a price-discovery problem, not a valuation signal. It is the same mechanic that produces triple-digit first-day pops on the STAR Market with some regularity — the difference here is the absolute size of the company being repriced.

    Is a $66 billion valuation defensible?

    Not on current financials. Unitree reported 2025 revenue of 1.7 billion yuan ($252 million) and net profit that Forbes put at 278 million yuan (about $41 million). At the closing price that is roughly 210 times sales and a price-to-earnings ratio near 1,300x.

    At the opening print, the revenue multiple was closer to 262x by Invezz’s calculation.

    For context, Nvidia at the height of its 2024 run traded at a fraction of that sales multiple while growing far faster off a vastly larger base.

    What the shipment data actually shows

    Unitree shipped about 5,500 humanoid units in 2025 and roughly 18,000 cumulatively through July 2026, according to Forbes. That is real product moving — more than most Western competitors can claim — but it is a rounding error against a $48–66 billion market cap.

    The margin trend is the harder problem. Invezz reported that adjusted net profit fell more than 52% year over year in the first quarter even as revenue grew 68%, as R&D and sales spending climbed. Growth is being bought, not compounded.

    Note also that profit figures diverge across outlets — Fortune cited a materially higher 2025 net profit of 600 million yuan ($89 million). Investors pricing a stock at four figures of earnings should probably know which number is right.

    Who profits from the Unitree listing?

    The pre-IPO cap table, overwhelmingly. Founder and chairman Wang Xingxing holds roughly 121.4 million shares, worth about 103 billion yuan at the close, per SCMP — a paper fortune built in under a decade.

    • Meituan holds an 8.7% stake worth roughly 30 billion yuan, which SCMP calculated as about a 70x return on its early investment.
    • Retail lottery winners captured a one-day gain most funds will not see in a decade.
    • Late buyers paid up to 1,100 yuan for a company that closed at 845 — a 23% loss inside a single session.
    • The STAR Market itself gets a marquee listing at a moment when Beijing wants domestic capital funding domestic hard tech.

    The broader tape was less enthusiastic. On the same day Unitree debuted, the STAR Market Composite Index fell 7.2% and the Shanghai Composite dropped 2.4%, SCMP reported. Money did not flow into robotics — it rotated out of everything else and into one ticker.

    What are the biggest risks to Unitree?

    Policy and adoption, in that order. Fortune reported that about 45% of Unitree’s sales are international, with the United States contributing 18% of 2025 revenue — exposure that a US robot import ban would hit directly.

    The demand case is also unproven at scale. HSBC researchers told Fortune that without major AI model improvements, “the current humanoid robot shipment upcycle is unlikely to be sustained over the next 1-2 years.”

    Nomura took the other side, crediting Unitree’s “rapid product iteration and continuous innovation” for a genuine first-mover advantage.

    Unitree itself has flagged that slower uptake of general-purpose robots could weigh on growth. When the issuer is the most cautious voice in the room, that is worth noting.

    Why this matters for the wider AI market

    Embodied AI just got its first public comparable, and it printed at a number nobody in the private market can match. Fortune noted that Unitree’s peak valuation exceeded Figure AI’s $39 billion mark from September 2025, making it the world’s most valuable robotics company.

    Forbes put the gap even more starkly: Agility Robotics is valued near $4 billion via SPAC merger — roughly 13 times smaller — despite more than $300 million in committed multi-year orders.

    That reprices every private robotics round still to come. Founders will point at Shanghai; investors will point at the fundamentals. Expect that argument in every Series B pitch this quarter.

    It also fits a pattern this blog has tracked all month: capital is chasing the physical layer of AI, not just the model layer. The same impulse drove SMIC’s first $3 billion quarter and the $1.1 billion raised by two-month-old River AI. Compute and hardware are where the money is going.

    And it sharpens the question hanging over every large AI private company — from Cognition at $40 billion to Anthropic’s reported $2 trillion IPO ambitions: what happens when a public market with limited float meets a private valuation built on projections?

    This post is reporting and analysis, not financial advice.

    Frequently asked questions about the Unitree IPO

    How much did Unitree raise in its IPO?

    Unitree raised 6.1 billion yuan, roughly $904 million, selling 40.45 million shares at 150.80 yuan each — about 10% of its enlarged share capital, according to Bloomberg.

    What is Unitree worth after the IPO?

    It touched about 444.9 billion yuan ($66 billion) at the open and closed near 342 billion yuan (roughly $48 billion), per SCMP. Sources vary between $48 billion and $53 billion for the close.

    Where does Unitree trade?

    On the Shanghai Stock Exchange’s STAR Market, China’s Nasdaq-style board for hard-tech companies. It is the first listed pure humanoid robot maker.

    Is Unitree profitable?

    Yes, but thinly. Forbes reported 2025 net profit of 278 million yuan ($41 million) on 1.7 billion yuan of revenue. Fortune cited a higher 600 million yuan figure. Q1 adjusted profit fell over 52% year over year.

    How many robots has Unitree sold?

    About 5,500 humanoid units in 2025 and roughly 18,000 cumulatively through July 2026, per Forbes.

    Who owns Unitree?

    Founder Wang Xingxing controls roughly 30% directly and indirectly, per The Standard. Meituan holds 8.7%, a stake SCMP valued at about 30 billion yuan.

    Can foreign investors buy Unitree shares?

    Access to STAR Market shares is restricted for most foreign retail investors and typically requires qualified institutional channels or Stock Connect eligibility, which varies by listing.

    The bottom line

    The Unitree IPO priced a scarcity, not a business. A 10% float, a 5,500x oversubscribed retail tranche and zero listed comparables produced a number that no earnings model supports.

    That does not make Unitree a bad company. It ships more humanoids than anyone, it is profitable, and it has a real first-mover position in a market that could be enormous.

    It makes the price a bet on 2030 revenue being 50 times 2025 revenue, with margins that are currently going the wrong way.

    Watch two things from here. First, whether the float expands after lockups and how the stock absorbs it. Second, whether US import restrictions bite into that 18% of revenue. Either would test a valuation with, as Invezz put it, “little room for operational disappointment.”

    The more consequential effect may be elsewhere. Every private robotics company now has a public number to anchor to — and every institutional investor now has a multiple to argue against.

    Sources

  • Claude Code vs Codex CLI: Which Coding Agent Wins on Cost

    Claude Code vs Codex CLI comes down to token math, not vibes. Codex bills GPT-5.6 Sol at $5/$30 per million tokens. Claude Opus 5 runs $5/$25 — but Anthropic’s own docs warn its newer tokenizer emits roughly 30% more tokens for the same text. That erases the discount. Codex wins on cost per solved task. Claude Code wins on agent orchestration.

    Both agents shipped meaningful releases this month, and both are now close enough on quality that price is the deciding variable. Here is the full breakdown, with every figure sourced.

    What changed for Claude Code vs Codex CLI in the last week?

    Anthropic shipped four Claude Code builds between August 13 and August 18, 2026. The headline change is subagent forking on by default. OpenAI’s Codex CLI moved earlier in the month, adding a one-command importer that pulls your Claude Code setup straight into Codex.

    Claude Code v2.1.232 made subagent_type: "fork" the default, so spawned agents inherit the full conversation and prompt cache. It also added cross-session @ mentions, letting one session message another by name.

    v2.1.233 added GitLab merge request support on August 15. v2.1.234, first seen August 18, added automatic session continuation when you hit a usage limit — a direct answer to the biggest complaint about subscription-metered agents.

    Codex CLI v0.147.0, dated August 4, 2026, added --approve-for-me for auto-reviewed approvals, opt-in support for the MCP 2026-07-28 protocol, and /import, which pulls supported setup and recent chats from Claude Code and Cursor. OpenAI is openly courting defectors.

    Which is better for coding, Claude Code or Codex CLI?

    On raw capability the two are inside each other’s error bars. The official Terminal-Bench harness leaderboard puts Claude Code first by 0.7 points. Artificial Analysis, running the same benchmark against models rather than harnesses, puts OpenAI first by 0.4 points. Neither gap is decisive.

    The harness leaderboard gives Claude Code the edge

    On the official Terminal-Bench 2.1 leaderboard, Claude Code running Fable 5 sits at 83.8% (±1.2), submitted June 7, 2026. Codex with GPT-5.5 follows at 83.1% (±1.1).

    Below that, Cursor CLI on Grok 4.5 scores 79.3%, Claude Code on Opus 4.8 hits 78.9%, and Codex on GPT-5.6 Terra lands at 78.4%. Gemini CLI on Gemini 3.1 Pro trails at 65.8%.

    Read that spread carefully: the harness matters less than the model inside it. Swapping Fable 5 for Opus 4.8 inside the same Claude Code wrapper costs 4.9 points.

    The model leaderboard gives OpenAI the edge

    Artificial Analysis runs Terminal-Bench v2.1 across 89 curated tasks spanning software engineering, sysadmin, data processing, model training and security. GPT-5.6 Sol at xhigh effort tops it at 89.5%.

    Claude Opus 5 at max effort scores 89.1%. Grok 4.6 at high effort takes third with 88.4%. A 0.4-point gap on an 89-task benchmark is roughly one third of a task. Call it a tie.

    How much do Claude Code and Codex CLI cost in 2026?

    Anthropic prices Claude Opus 5 at $5 input and $25 output per million tokens. OpenAI prices GPT-5.6 Sol at $5 and $30. On the sticker, Anthropic is 17% cheaper on output. Subscription tiers land within $20 of each other at every level.

    Token prices, side by side

    ModelInput / 1MOutput / 1MCached inputContext
    Claude Opus 5$5$25$0.50200K (1M available)
    Claude Sonnet 5$2$10$0.20200K
    Claude Fable 5$10$50$1200K
    GPT-5.6 Sol$5$30$0.501.05M
    GPT-5.6 Terra$2$12$0.201.05M
    GPT-5.6 Luna$0.20$1.20$0.021.05M
    Sources: Claude Platform pricing docs; OpenAI GPT-5.6 announcement (July 30, 2026) and published API rate card, August 18, 2026 snapshot.

    The tier structures mirror each other almost exactly. Sonnet 5 and Terra both cost $2 input. Anthropic has no answer to Luna’s $0.20/$1.20, which OpenAI claims beats Fable 5 on Agents’ Last Exam at an estimated cost per task nearly 99% lower.

    Subscription plans and what they actually buy

    • Claude Free — $0. Claude Code included, limits reset on a rolling five-hour window.
    • Claude Pro — $20/month, or $17/month billed annually. At least 5x Free usage per session.
    • Claude Max — from $100/month, in 5x and 20x tiers above Pro.
    • Codex Go — $8/month, web and CLI access only.
    • Codex Plus — $20/month, adds IDE and iOS.
    • Codex Pro — $100/month (5x) and $200/month (20x).

    One structural difference matters. Anthropic pools your usage: web, desktop, mobile and Claude Code all draw from the same bucket. Chat with Claude all afternoon and your coding agent gets less runway.

    OpenAI meters Codex in credits. Its published rate card charges 125 credits per million Sol input tokens and 750 per million output. At the standard $0.04 per credit, that is exactly the $5/$30 API rate — no subscription markup, no discount. OpenAI’s own guidance puts real-world spend at roughly $100 to $200 per developer per month.

    What is the real cost per solved task?

    Price per token is the wrong metric. What you pay is price per token multiplied by tokens consumed, and on that measure Codex currently wins by a wide margin. Two multipliers work against Claude: a heavier tokenizer and a chattier agent loop.

    The tokenizer tax nobody prices in

    Anthropic’s own pricing documentation carries a warning most buyers skip: Claude 4.7 and later use a newer tokenizer that produces about 30% more tokens for the same text than earlier models.

    Run the arithmetic. Opus 5 output at $25 per million, inflated 30%, is an effective $32.50 for the text volume that Sol bills at $30. Anthropic’s 17% headline discount becomes an 8% premium.

    The token-efficiency gap is the bigger number

    Independent head-to-head testing published by Morph puts the consumption gap at 3.2x to 4.2x. On a Figma plugin build, Codex burned 1,499,455 tokens against Claude Code’s 6,232,242. On a scheduler app, 72,579 against 234,772.

    Combine the two effects and the verdict is brutal. A 17% per-token discount against a 4.2x volume gap leaves Claude Code roughly 3.5x more expensive per completed task in that test set.

    That gap is a workflow choice, not a defect. Claude Code explores more, forks more subagents, and reads more of your repo. You are paying for breadth. Whether breadth is worth 3.5x depends entirely on whether your tasks are ambiguous or well specified.

    Which coding agent should you buy in 2026?

    Pick Codex if your work is well-specified and high-volume. Pick Claude Code if your work is exploratory, multi-repo, or heavy on parallel agents. Pick Luna or Sonnet 5 for anything routine — the flagship tier is wasted on boilerplate, and that is where most teams overspend.

    Your situationBuyMonthlyWhy
    Solo dev, tight budgetCodex Go$8Cheapest real CLI agent; token efficiency stretches it furthest
    Solo dev, daily driverClaude Pro (annual)$17Pooled usage across chat and code; cheapest flagship access
    High-volume refactorsCodex Plus or Pro$20–$2003–4x fewer tokens per task at near-identical benchmark scores
    Parallel agent swarmsClaude Max 20x$200Fork-by-default subagents, no published parallel-agent cap
    Large-context monorepoCodex on SolUsage-based1.05M-token context at flat pricing under the long-context threshold
    Routine boilerplateCodex on LunaUsage-based$0.20/$1.20 per million; 96% cheaper output than Sol
    Recommendations based on August 2026 published pricing and benchmark data.

    Is Claude Code worth $200 a month in 2026?

    Yes, if you run agents in parallel. Max 20x buys 20x the per-session allowance of Pro, and Claude Code’s fork-by-default subagents inherit the prompt cache, so parallelism is cheap in wall-clock terms. Codex reportedly caps developers at eight parallel agents.

    No, if you are a single developer running one session at a time. At that duty cycle you will not exhaust Pro’s $17 annual tier, and the extra $183 buys headroom you never touch.

    The August 18 addition of automatic session continuation at usage limits changes the calculus slightly. Hitting a cap now costs you a pause rather than a lost session, which makes the cheaper tiers more tolerable.

    For the model layer underneath these harnesses, our breakdowns of Muse Spark vs Claude Opus 5 and Gemini 3.7 Flash vs Claude Sonnet 5 cover the cost-per-coding-point math in detail.

    Frequently asked questions

    Can I switch from Claude Code to Codex CLI without redoing my setup?

    Yes. Codex CLI v0.147.0, released August 4, 2026, added /import, which pulls supported setup files and recent chats from both Claude Code and Cursor.

    Which has the larger context window?

    Codex. All three GPT-5.6 tiers carry a 1.05M-token window. Anthropic’s pricing page lists 200K as standard, noting that 4.5+ models include a 1M-token window at standard pricing.

    Does Claude Code cost extra on top of a Claude subscription?

    No. Claude Code is included on Free, Pro and Max, and shares the same usage pool as web, desktop and mobile. You can enable pay-as-you-go credits on paid plans once you hit a limit.

    Is the cheapest model good enough for coding?

    For routine work, often yes. OpenAI states Luna outperforms Fable 5 on Agents’ Last Exam at an estimated cost per task nearly 99% lower. On Terminal-Bench 2.1, Codex on Luna still scored 75.7%.

    Which agent tops the Terminal-Bench leaderboard right now?

    Claude Code with Fable 5, at 83.8%, leads the official harness leaderboard. On Artificial Analysis’s model-level run of the same benchmark, GPT-5.6 Sol leads at 89.5%.

    Do open-weight models beat either of these on cost?

    On price per token, easily. On benchmark parity, not yet at the top tier. We compared the leading options in GLM-5.3 vs DeepSeek V4 Pro.

    Will these prices hold?

    Assume not. The DeepSeek price increase showed how fast the floor can move. OpenAI cut Luna’s rate on July 30 and raised long-context rates at the same time.

    The bottom line

    Buy Codex if you are optimizing spend. The benchmark gap between the two is smaller than the measurement error, and Codex delivers that parity at roughly a third to a quarter of the token consumption. That is not a rounding difference — on a $200/month budget it is the gap between running out on day 20 and finishing the month.

    Buy Claude Code if you are optimizing throughput on ambiguous work. Fork-by-default subagents, cross-session mentions and an uncapped parallel-agent model make it the better orchestrator, and orchestration is where the next year of agent productivity gets won.

    What it depends on, precisely: if your average task ships with a clear spec and acceptance criteria, Codex is cheaper for identical output. If your average task starts with “figure out why this breaks,” Claude Code’s extra tokens are buying you something real.

    The move most teams should make is neither: run Luna or Sonnet 5 for 80% of tasks and reserve the flagship tier for the 20% that need it. That single routing decision saves more money than picking the right vendor.

    Sources

  • River AI Funding Round: $1.1B for a Two-Month-Old Startup

    The River AI funding round closed at $1.1 billion across seed and Series A, led by General Catalyst and AMP PBC, with Nvidia, AMD Ventures, Y Combinator and Temasek participating. The company, founded by xAI co-founder Igor Babuschkin, is roughly two months old and has not disclosed a valuation. It sells fine-tuning and reinforcement learning for open-weight models.

    How much did River AI raise, and from whom?

    River AI raised $1.1 billion in a combined Series Seed and Series A, according to the company’s August 11 announcement distributed via Business Wire. General Catalyst and AMP PBC led. Nvidia and AMD Ventures came in as strategic investors, alongside Y Combinator and Temasek.

    No post-money valuation was disclosed. That omission is the single most important detail in the entire River AI funding round, and we will come back to it.

    The company is headquartered in Palo Alto. TechCrunch reported that River AI was roughly two months old at the time of the raise, having launched in June 2026.

    Deal terms at a glance

    Item Detail
    Total raised $1.1 billion
    Structure Series Seed + Series A, announced together
    Lead investors General Catalyst, AMP PBC
    Strategic investors Nvidia, AMD Ventures
    Other investors Y Combinator, Temasek
    Post-money valuation Not disclosed
    Announcement date August 11, 2026
    Company age ~2 months (launched June 2026)
    Headquarters Palo Alto, California
    Founder and CEO Igor Babuschkin, xAI co-founder

    Who is Igor Babuschkin, and why does his name move this much money?

    Babuschkin is the reason a two-month-old company cleared ten figures. He co-founded xAI. Before that he worked at Google DeepMind, where he contributed to AlphaCode, the first coding model to place competitively in a programming contest. He also spent time at OpenAI.

    He left xAI in August 2025. CNBC reported at the time that he was departing to start an AI-safety-focused venture capital firm. Twelve months later he is running an AI infrastructure company instead.

    That pivot deserves more scrutiny than it has received. The stated plan was to allocate capital to safety research. The executed plan was to raise $1.1 billion and build a training platform. Investors appear untroubled by the change.

    What does River AI actually sell?

    River AI sells the River API: reinforcement learning and LoRA fine-tuning for open-weight models, billed per million tokens. SiliconANGLE reported the platform supports models from 35 billion to 1 trillion parameters.

    LoRA — low-rank adaptation — adds a small set of trainable parameters on top of a frozen base model. It is cheap relative to full retraining. It is also not novel; it is standard practice across the open-weight ecosystem.

    The company’s differentiation claims are about speed and cost:

    • Complex reinforcement learning runs completed in 15 to 20 minutes, per the company’s own press release
    • No dedicated infrastructure team required on the customer side
    • Customized models up to four times more cost-efficient than proprietary alternatives, per company claims
    • Instant deployment to production with token-metered billing

    Every one of those figures is a vendor claim. None has been independently benchmarked. Treat them accordingly.

    The silicon ambition

    River AI also intends to build custom chips. SiliconANGLE reported plans for machine learning accelerators on advanced foundry nodes, plus a PyTorch compiler to run models efficiently on that proprietary hardware.

    This is where the $1.1 billion starts to make arithmetic sense. A fine-tuning API does not need a billion dollars. A custom accelerator program does — and still probably needs more.

    Why did Nvidia and AMD both back the same startup?

    Nvidia and AMD Ventures are on the same cap table. These are direct competitors in AI accelerators, and River AI has publicly stated it wants to build competing silicon.

    The read: both are buying optionality, not conviction. Strategic investment at this stage is cheap intelligence on a team that could matter later. It is also a hedge against a customer base that increasingly wants open-weight models running on non-Nvidia hardware.

    General Catalyst’s Hemant Taneja framed it in national terms, saying the firm views River AI’s agenda “as a priority for American resilience.” That is a strategic-narrative sale, not a unit-economics sale.

    Nvidia has spent 2026 turning its balance sheet into an instrument of demand creation — a pattern we examined in our coverage of Nvidia’s revised OpenAI data center guarantee. A small strategic check into an open-weight training platform fits that playbook precisely.

    How does this compare to other record AI seed rounds?

    The benchmark is Thinking Machines Lab. TechCrunch reported in July 2025 that Mira Murati’s startup raised a $2 billion seed at a $12 billion valuation — the largest seed round on record.

    River AI’s $1.1 billion is smaller in absolute terms but sits in the same tier of pre-product capital formation. The difference is disclosure: Thinking Machines named its valuation. River AI did not.

    Company Round Amount Valuation Date
    Thinking Machines Lab Seed $2.0B $12B Jul 2025
    River AI Seed + Series A $1.1B Not disclosed Aug 2026
    Lovable Series C $400M $13.3B Aug 2026
    Databricks Strategic $5.0B See coverage Aug 2026

    Context on the week itself: StartupHub.ai counted roughly $10 billion in disclosed AI capital across about 40 rounds between August 11 and August 17. River AI alone was more than a tenth of it.

    The same week, Lovable confirmed a $400 million Series C at a $13.3 billion valuation, per TechCrunch and Bloomberg — a valuation that doubled in roughly seven months. And Databricks took $5 billion in a strategic round led by Coatue and Blackstone, extending the trajectory we covered in Databricks’ $190 billion valuation.

    Why this matters

    Three things follow from the River AI funding round.

    First, open-weight infrastructure is now a fundable category on its own. The bet is that enterprises will run and customize open models rather than rent frontier APIs — the economics we broke down in our analysis of Qwen3.8-Max open weights versus API pricing.

    Second, founder pedigree is being priced as an asset class. Two months of operating history and no disclosed valuation did not slow this round. Compare that with Cognition’s $40 billion valuation, which at least came with shipped products and revenue.

    Third, the toll-booth layer of AI keeps attracting capital. Routing, fine-tuning and metering are where margin is accumulating — the same logic behind Stripe’s $7 billion OpenRouter acquisition.

    For investors, the honest summary is that this is a pre-revenue bet on a person and a thesis. This post is reporting and analysis, not financial advice.

    What should skeptics watch?

    Watch the valuation. A round announced without one usually means the number is either uncomfortable to defend or structured with terms that complicate the headline. Neither is disqualifying. Both are worth knowing.

    Watch the benchmarks. The 15-to-20-minute training claim and the four-times cost advantage are unverified vendor figures. Independent replication would change the story considerably.

    Watch the silicon timeline. Custom accelerators on advanced foundry nodes take years and consume capital faster than any API can generate it. $1.1 billion is a down payment, not a war chest.

    Frequently asked questions

    How much did River AI raise?

    $1.1 billion, announced August 11, 2026, structured as a combined Series Seed and Series A.

    What is River AI’s valuation?

    Not disclosed. Neither the company’s press release nor coverage from TechCrunch or SiliconANGLE reported a post-money figure.

    Who led the River AI funding round?

    General Catalyst and AMP PBC led. Nvidia and AMD Ventures joined as strategic investors, with Y Combinator and Temasek also participating.

    Who founded River AI?

    Igor Babuschkin, a co-founder of xAI who previously worked at Google DeepMind on AlphaCode and at OpenAI. He left xAI in August 2025.

    What does River AI sell?

    The River API: reinforcement learning and LoRA fine-tuning for open-weight models between 35 billion and 1 trillion parameters, billed per million tokens.

    Is River AI competing with Nvidia?

    Eventually, yes. The company has said it plans custom machine learning accelerators and a PyTorch compiler — while Nvidia sits on its cap table.

    How does this compare to the largest AI seed round ever?

    Thinking Machines Lab raised $2 billion at a $12 billion valuation in July 2025, per TechCrunch. River AI’s $1.1 billion is smaller but in the same tier.

    The bottom line

    The River AI funding round is a $1.1 billion wager that open-weight customization becomes infrastructure, placed on a founder with a genuine frontier-lab record and a two-month-old company.

    The capital is real and the investor list is serious. The product claims are not yet independently verified, and the missing valuation is a gap that will get filled — one way or another — at the next round.

    Expect two signals over the next two quarters: a disclosed valuation, and third-party benchmarks against proprietary fine-tuning services. If both land well, this looks early. If neither does, this looks like 2026’s clearest example of pedigree pricing outrunning product.

    Sources

  • GLM-5.3 vs DeepSeek V4 Pro: Which Open-Weight Coder to Buy

    GLM-5.3 vs DeepSeek V4 Pro comes down to one thing: you can download DeepSeek today. Its MIT-licensed weights shipped August 12, 2026, with a 1M-token context, and DeepSeek’s own API lists $0.435 per million input tokens against $0.87 output. GLM-5.3 landed two days later with stronger cyber scores, no published per-token price, and weights held back roughly two weeks.

    Two of the most consequential open-weight coding models of the year shipped 48 hours apart. Both vendors published benchmark tables claiming frontier-class performance. Neither number has been independently replicated.

    Here is what the money actually says.

    What shipped in the GLM-5.3 vs DeepSeek V4 Pro week?

    DeepSeek V4 Pro 0813 went generally available on August 12, 2026, after a preview build dated April 24. Z.ai launched GLM-5.3 on August 14. Both target agentic coding and terminal work. Only one of them can be run on your own hardware right now.

    DeepSeek V4 Pro 0813: 1.6 trillion parameters, MIT license

    V4 Pro is a mixture-of-experts model with 1.6 trillion total parameters and 49 billion active per token, according to TechTimes’ launch coverage. Context window is 1,048,576 tokens. The license is MIT — the most permissive terms in the frontier tier.

    Artificial Analysis ranks it third out of 1,075 models evaluated on its Intelligence Index, with a score of 53, and measures output at 75.4 tokens per second with a 1.69-second time to first token.

    GLM-5.3: same base model, rebuilt post-training

    Z.ai did something unusual. GLM-5.3 reuses the 743-billion-parameter base model from GLM-5.2 with no new pre-training run, per MarkTechPost’s technical breakdown. Every gain came from post-training.

    The gains are not small. Terminal-Bench 3.0 went from 4.6 to 28.3. DeepSWE v1.1 climbed from 46.2 to 66.9. That is a roughly 45% relative jump on agentic software engineering from post-training alone.

    Which is better for coding, GLM-5.3 or DeepSeek V4 Pro?

    On raw self-reported coding scores, DeepSeek V4 Pro leads. It claims 62.7 on DeepSWE and 87.9 on Terminal-Bench 2.1. GLM-5.3 reports 66.9 on DeepSWE v1.1 and 28.3 on Terminal-Bench 3.0. The benchmarks are different versions, so the comparison is not apples to apples — and that is the whole problem.

    The benchmark numbers nobody has verified

    DeepSeek’s self-reported 87.9 on Terminal-Bench 2.1 would place it first in the world. It does not appear on the benchmark’s own board.

    The official Terminal-Bench 2.1 leaderboard is topped by Claude Code running Fable 5 at 83.8% ± 1.2%, followed by Codex with GPT-5.5 at 83.1% and Cursor CLI with Grok 4.5 at 79.3%. No DeepSeek entry exists.

    TechTimes noted the same gap: V4 Pro’s scores “have not been independently replicated by any third-party evaluator as of publication.”

    Independent evaluator Artificial Analysis runs its own Terminal-Bench v2.1 harness and gets different absolute numbers again: GPT-5.6 Sol at xhigh effort scores 89.5%, Claude Opus 5 at max effort 89.1%, Grok 4.6 at 88.4%.

    Three sources, three scales, zero cross-comparability. Treat every vendor coding number as a marketing claim until a third party runs the harness.

    Where GLM-5.3 actually wins

    Security. GLM-5.3 scores 84.5% on CyberGym, up from 77.2%, ahead of Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%. DeepSeek V4 Pro self-reports 83.3%.

    ExploitBench more than doubled, from 24.4% to 54.4%. Z.ai says the model surfaced 2,436 vulnerabilities across 269 open-source projects, 1,097 of them rated critical or high severity.

    If your workload is vulnerability triage or defensive security automation, that is the strongest open-weight number on the board.

    How much does each model cost per million tokens?

    DeepSeek is cheaper — but only if you buy direct. The same model costs roughly 2.8x more input and 2.8x more output through a marketplace. GLM-5.3 has no published per-token rate at all; Z.ai’s price table still lists GLM-5.2. That pricing opacity is a real cost.

    Metric DeepSeek V4 Pro 0813 GLM-5.3
    Launch date Aug 12, 2026 (GA) Aug 14, 2026
    Input / 1M (vendor direct) $0.435 cache miss Not published (GLM-5.2: $1.40)
    Cached input / 1M $0.003625 Not published (GLM-5.2: $0.26)
    Output / 1M (vendor direct) $0.87 Not published (GLM-5.2: $4.40)
    Input / output via OpenRouter $1.218 / $2.436 GLM-5.2: $0.50 / $3.15
    Parameters 1.6T total / 49B active 743B base (shared with 5.2)
    Context window 1,048,576 tokens Not disclosed
    License MIT, weights live Weights ~2 weeks post-launch
    Subscription option None $18 / $80 / $168 per month

    Sources: TechTimes, OpenRouter, OpenRouter GLM 5.2.

    The marketplace markup is the hidden tax

    Three sources quote three different prices for the identical DeepSeek model. TechTimes lists DeepSeek’s own $0.435 / $0.87. OpenRouter lists $1.218 / $2.436. Artificial Analysis measures $1.32 input and $3.96 output, with a blended rate of $0.69 per million at a 7:2:1 ratio.

    On a 500-million-token month at a typical 3:1 input-output split, that spread is real money: roughly $272 buying direct against roughly $685 through OpenRouter. Same weights, same model ID, 2.5x the invoice.

    DeepSeek has also signaled another increase is coming, with no timeline. Anyone budgeting off today’s rate should read our breakdown of the DeepSeek price increase that ended the AI price war.

    Subscription versus per-token

    Z.ai’s answer to pricing volatility is a flat plan. The GLM Coding Plan runs $18/month for Lite, $80 for Pro and $168 for Max, with weekly credit quotas of 10,000, 60,000 and 140,000 respectively.

    For a solo developer hammering an agent all day, $18 flat beats metered billing on predictability alone. For an API-backed product with variable load, per-token wins on unit economics.

    Can you actually download the weights?

    DeepSeek yes, GLM-5.3 not yet. This is the single most decisive difference between the two models and it gets buried under benchmark tables. An open-weight model you cannot download is a closed model with a press release.

    Z.ai says GLM-5.3 weights land roughly two weeks after the August 14 launch, once “safety evaluation and hardening” finish. Until then, access is API, GLM Coding Plan, or ZCode only.

    That delay is defensible given the model’s exploit-generation scores. It is still a delay, and it has three concrete consequences:

    • No air-gapped deployment. Regulated teams that cannot send code to a Chinese API endpoint are locked out entirely.
    • No cost floor. You cannot benchmark self-hosted cost per token against the API rate, which is the entire argument for open weights.
    • No fork risk protection. If Z.ai changes pricing or terms, there is no downloaded checkpoint to fall back on.

    DeepSeek’s MIT license carries none of those problems. For the economics of running weights yourself, see our analysis of what self-hosting Qwen3.8-Max really costs.

    Which model should you pick for your use case?

    Pick DeepSeek V4 Pro for long-context work, self-hosting and cost-sensitive production. Pick GLM-5.3 for security workloads and for flat-rate interactive coding. Neither beats Claude Fable 5 or GPT-5.5 on the independently verified Terminal-Bench leaderboard, so neither is the right call if raw capability is your only constraint.

    Use case Pick Why
    Self-hosted / air-gapped DeepSeek V4 Pro MIT weights available now
    Large-repo refactors DeepSeek V4 Pro 1M-token context confirmed
    High-volume API product DeepSeek V4 Pro $0.435 / $0.87 direct rate
    Vulnerability triage GLM-5.3 84.5% CyberGym, top open score
    Solo dev, predictable bill GLM-5.3 $18/month Coding Plan floor
    Long-horizon agent runs GLM-5.3 Terminal-Bench 3.0: 4.6 to 28.3
    Absolute best coding score Neither Fable 5 leads at 83.8% verified

    Is GLM-5.3 worth it in 2026?

    Yes, but narrowly. GLM-5.3 is worth paying for if you are doing security work or want a fixed monthly bill. It is not worth waiting for if you need weights on your own GPUs this quarter, and it is not the best coder available at any price.

    Z.ai calls it “the strongest open-weights coder on the market.” Its own numbers complicate that. On Terminal-Bench 3.0, GLM-5.3’s 28.3 trails Claude Fable 5 at 33.7 and GPT-5.6 Sol at 34.6. On DeepSWE v1.1, its 66.9 trails Kimi K3 at 67.5 and Fable 5 at 69.7.

    The one benchmark where GLM-5.3 leads outright is Z.ai Code Bench — Z.ai’s own benchmark. It scores 31.4% at roughly 50,000 tokens per task against Claude Opus 4.8 at 29.5% using 120,000 tokens. Efficient, and self-graded.

    Efficiency is the real story there. Beating a frontier closed model while spending 58% fewer tokens per task is a genuine margin advantage — if the benchmark holds up externally.

    Frequently asked questions

    Is DeepSeek V4 Pro open source?

    The weights are MIT-licensed and available, which permits commercial use, modification and redistribution. Training data and code are not released, so it is open-weight rather than fully open-source.

    When do GLM-5.3 weights release?

    Z.ai said roughly two weeks after the August 14, 2026 launch, pending safety evaluation and hardening. That points to late August. No license has been confirmed.

    What is the cheapest way to run DeepSeek V4 Pro?

    Direct through DeepSeek’s API at $0.435 per million input tokens on a cache miss and $0.87 output. Cache hits drop input to $0.003625. Marketplace routing costs roughly 2.5x more.

    Does GLM-5.3 beat Claude on coding?

    No, on the numbers Z.ai published. GLM-5.3 scores 28.3 on Terminal-Bench 3.0 against Claude Fable 5’s 33.7. It leads only on Z.ai’s internal Code Bench and on CyberGym.

    Why do Terminal-Bench scores differ between sources?

    Because the harness and agent scaffold change the result. Claude Code with Fable 5 scores 83.8% on the official 2.1 board, while Artificial Analysis’s own v2.1 run puts GPT-5.6 Sol at 89.5%. Only compare scores measured by the same evaluator.

    Which model has the bigger context window?

    DeepSeek V4 Pro, at 1,048,576 tokens confirmed by both OpenRouter and Artificial Analysis. Z.ai has not disclosed GLM-5.3’s context length.

    Is GLM-5.3’s cyber capability a risk?

    Z.ai treated it as one, holding weights for safety hardening after ExploitBench scores doubled to 54.4%. The model reportedly found 1,097 critical or high-severity vulnerabilities across 269 open-source projects.

    The bottom line

    Buy DeepSeek V4 Pro. Today, direct from DeepSeek, at $0.435 in and $0.87 out.

    It has the weights you can actually download under MIT, the 1M-token context you can actually verify, and a price roughly 3x below what the same model costs through a marketplace. Third place out of 1,075 models on Artificial Analysis’s Intelligence Index is enough capability for the overwhelming majority of production coding work.

    Buy GLM-5.3 in exactly two situations: your workload is security-focused, where 84.5% on CyberGym is the best open-weight number published; or you want a $18-to-$168 monthly ceiling instead of metered billing that DeepSeek has already warned will rise again.

    And discount both vendors’ headline coding claims. DeepSeek’s 87.9 on Terminal-Bench 2.1 would top the world leaderboard, and DeepSeek is not on that leaderboard. Until a third party runs the harness, those are sales figures, not results. For a comparison where the numbers were independently checked, see our breakdown of cost per coding point across Gemini 3.7 Flash and Claude Sonnet 5, and our look at Meta’s 4x cheaper coding model.

    Sources

  • Stripe OpenRouter Acquisition: $7 Billion for a 5% Toll on AI Tokens

    The Stripe OpenRouter acquisition closes at more than $7 billion, Bloomberg reported on August 16, 2026 — over five times the $1.3 billion valuation OpenRouter carried in May, when it raised a $113 million Series B. Stripe is buying a routing layer that moves roughly 25 trillion tokens a week for 8 million developers, and takes a cut of every one of them.

    What is the Stripe OpenRouter acquisition?

    Stripe has agreed to buy OpenRouter, the gateway that lets developers call 400-plus AI models through a single API, for more than $7 billion. Bloomberg first reported the finalized agreement on August 16. Fortune confirmed the figure the same day. Neither company would comment.

    OpenRouter was founded in 2023 and is based in New York. Its CEO, Alex Atallah, co-founded the NFT marketplace OpenSea. He has described OpenRouter as “the AI equivalent of Stripe” — a line that reads differently now.

    Per Fortune, the company had raised over $150 million in total before the deal.

    The deal terms at a glance

    Item Figure Source
    Reported purchase price More than $7 billion Bloomberg, Aug 16, 2026
    Earlier reported price ~$10 billion WSJ, July 2026
    Valuation, May 2026 $1.3 billion Series B announcement
    Series B size / lead $113 million / CapitalG SiliconANGLE, May 26, 2026
    Total capital raised pre-deal Over $150 million Fortune
    Weekly token throughput ~25 trillion SiliconANGLE, May 2026
    Developers on platform 8 million Fortune / TechCrunch
    Models available 400+ TechCrunch

    How the price fell from $10 billion

    The Wall Street Journal reported in July that Stripe was in talks at roughly $10 billion. The finalized number is about 30% below that. Fortune notes the final price remains subject to change.

    A 30% haircut between leak and signature is not nothing. Either diligence found something, or the July number was a seller’s anchor that never had a buyer behind it.

    What does OpenRouter actually do?

    OpenRouter is a single API endpoint that sits in front of hundreds of model providers. A developer writes one integration, then swaps between OpenAI, Anthropic, Google, DeepSeek, Alibaba and dozens of open-weight hosts by changing a string — no new contract, no new billing relationship.

    The pitch is that model choice is now a per-request decision, not a procurement decision.

    25 trillion tokens a week

    SiliconANGLE reported at the Series B that OpenRouter was routing about 25 trillion tokens per week, up from roughly 5 trillion six months earlier — a fivefold increase in half a year.

    That growth is the whole thesis. The company does not train models, does not own GPUs, and does not sell inference capacity. It sells the seam between all of them.

    Its Series B investor list reads like a map of who benefits from that seam existing:

    • CapitalG — Alphabet’s growth fund, lead investor
    • NVentures — Nvidia’s venture arm
    • Andreessen Horowitz, Menlo Ventures, Sequoia — earlier backers
    • ServiceNow, MongoDB, Snowflake and Databricks Ventures — enterprise data platforms with their own routing problems

    When four enterprise data vendors and two chip-adjacent funds all buy into the same routing layer, they are hedging the same risk: that a single model vendor captures the application tier. We covered a related version of that bet in our piece on Databricks at a $190 billion valuation.

    Why is Stripe paying $7 billion for a routing layer?

    Because Stripe is assembling the billing stack for usage-priced software, and inference is the largest new usage-priced category in the market. OpenRouter is not a payments company Stripe is absorbing. It is a meter Stripe now owns.

    The Metronome and Bridge pattern

    This is the third leg of a visible strategy. Stripe bought stablecoin platform Bridge for a reported $1.1 billion in 2025, then closed its acquisition of usage-based billing company Metronome on January 14, 2026. Metronome already handled metering for OpenAI, Anthropic and Nvidia, per Stripe’s own announcement.

    Stripe CEO Patrick Collison said at the time that “the shift toward usage-based models will be a defining feature of the next decade for our industry,” calling metering and billing “the interface between ‘product’ and ‘business.’”

    Metronome bills the tokens. OpenRouter routes them. Stripe now owns both ends of the same wire.

    The scale Stripe is bolting this onto

    Stripe’s 2025 annual letter reported $1.9 trillion in total payment volume, up 34% year over year, equal to roughly 1.6% of global GDP. A February 2026 tender offer valued the company at $159 billion, CNBC reported.

    At that size, $7 billion is about 4.4% of Stripe’s own valuation — expensive, but not existential. The company also said it remained “robustly profitable.”

    Is the $7 billion price justified?

    On revenue multiples, no — not obviously. OpenRouter does not publish financials, and the available estimates make the price look aggressive even by 2026 standards.

    The multiple problem

    Research firm Sacra estimates OpenRouter reached roughly $50 million in annualized revenue by March 2026, up from about $19 million at the end of 2025, on a commission of roughly 5% of inference spend. At $50 million, a $7 billion price is about 140x revenue.

    TechTimes reported annualized revenue closer to $140 million by mid-2026, which would put the deal near 50x. Both numbers are estimates. Neither is audited. Take them as a range, not a fact.

    There is also a structural oddity worth naming: OpenRouter’s take rate on inference spend is roughly 5%, while Stripe’s blended take rate on payment volume is a fraction of a percent. Stripe is paying a high multiple to acquire a much higher-margin toll — which only works if that toll survives contact with scale.

    The neutrality problem

    OpenRouter’s value proposition is that it is neutral. It picks the cheapest adequate model for a request, regardless of vendor. That neutrality is now owned by a company with its own billing interests and deep commercial ties to OpenAI, including the jointly developed Agentic Commerce Protocol.

    Nothing about the deal forces bias into the routing. But the incentive to keep the router perfectly indifferent is weaker on Monday than it was on Friday. Enterprise buyers will notice.

    The concentration problem

    OpenRouter’s traffic mix has shifted hard toward cheap Chinese open-weight models over the past year, according to platform data cited by TechTimes. That is good for volume and bad for the take rate, because 5% of a cheap token is less than 5% of an expensive one.

    It also puts a US payments company in the middle of a politically live supply chain — a House select committee opened an inquiry into Chinese model providers in April 2026. We wrote about the pricing pressure driving that shift in DeepSeek’s price increase and the end of the AI price war.

    Why this matters for the AI market

    The Stripe OpenRouter acquisition is a data point about where value is settling in the AI stack. It is not settling in the model.

    Three things follow:

    1. Infrastructure between models is repricing upward. A company with no models and no GPUs just cleared $7 billion. Compare that to Cognition’s $40 billion valuation — a product company — and the gap is narrowing on a revenue-multiple basis.
    2. Model commoditization is now an investable thesis. Routing is only worth $7 billion if buyers genuinely expect to switch models constantly. That is a bet against any single lab’s pricing power.
    3. Consolidation is accelerating. This is the third multi-billion-dollar AI acquisition in roughly a fortnight, alongside Anthropic’s $6 billion move for Decart and SpaceX’s $60 billion purchase of Cursor.

    For investors without access to private markets, the readthrough is indirect: Stripe is private, and the clearest public exposure is through Alphabet, whose CapitalG marked a roughly 5x return in three months. This post is reporting and analysis, not financial advice.

    Frequently asked questions

    How much did Stripe pay for OpenRouter?

    More than $7 billion, according to Bloomberg’s August 16, 2026 report. Fortune notes the final figure is subject to change. Neither company has confirmed it publicly.

    What was OpenRouter worth before the deal?

    $1.3 billion, set at its $113 million Series B in May 2026, led by Alphabet’s CapitalG. The acquisition price is more than five times that, roughly three months later.

    What does OpenRouter do?

    It provides one API that routes requests across 400-plus AI models from many providers, letting developers switch models on cost or capability without changing their integration or billing relationship.

    How much traffic does OpenRouter handle?

    About 25 trillion tokens per week as of May 2026, per SiliconANGLE — up from roughly 5 trillion tokens per week six months earlier.

    Why does a payments company want an AI router?

    Because inference is metered, and Stripe is building the metering stack. It closed the Metronome billing acquisition in January 2026; OpenRouter adds the routing layer that generates the meter readings.

    Is OpenRouter profitable?

    Unknown. The company does not publish financials. Sacra estimates roughly $50 million in annualized revenue as of March 2026 on a ~5% commission — an estimate, not a disclosure.

    Will OpenRouter stay neutral between model providers?

    Stripe has not said. The commercial logic of the acquisition depends on developers trusting the router to be indifferent, so any visible bias would damage the asset Stripe just bought.

    The bottom line

    Stripe paid a venture-scale multiple for an infrastructure position, not for a P&L. At more than $7 billion against estimated revenue somewhere between $50 million and $140 million, the price only makes sense if token volume keeps compounding and the 5% toll holds.

    Watch two things. First, whether the take rate survives as traffic migrates to cheap open-weight models — volume growth means nothing if the per-token cut collapses. Second, whether enterprise customers keep routing through a gateway owned by a company that also bills their competitors.

    The AI infrastructure land grab is no longer about chips and datacenters alone. It is about who owns the meter. Nothing here is financial advice.

    Sources